Stable Diffusion, NovelAI, Machine Learning Art - AI art generation discussion and image dump

  • 🐕 I am attempting to get the site runnning as fast as possible. If you are experiencing slow page load times, please report it.
tech moving fast

twitter_toyxyz(@toyxyz3)_20230201-044740_1620645033363017728_photo.jpg


twitter_toyxyz(@toyxyz3)_20230201-053249_1620656396487503872_photo.jpg



twitter_toyxyz(@toyxyz3)_20230202-165144_1621189637292294145_photo.jpg


twitter_toyxyz(@toyxyz3)_20230202-171638_1621195902693306368_photo.jpg


twitter_toyxyz(@toyxyz3)_20230207-070522_1622854013451395074_photo.jpg



twitter_toyxyz(@toyxyz3)_20230213-042741_1624988657307516928_photo.jpg



twitter_toyxyz(@toyxyz3)_20230213-043900_1624991506632445982_photo.jpg


twitter_toyxyz(@toyxyz3)_20230213-043901_1624991510801575937_photo.jpg



twitter_toyxyz(@toyxyz3)_20230214-152142_1625515633558564864_photo.jpg



twitter_toyxyz(@toyxyz3)_20230214-152222_1625515804027666433_photo.jpg



twitter_toyxyz(@toyxyz3)_20230214-152253_1625515930276225025_photo.jpg



twitter_toyxyz(@toyxyz3)_20230214-152321_1625516049448992768_photo.jpg



twitter_toyxyz(@toyxyz3)_20230214-152347_1625516157141921793_photo.jpg



twitter_toyxyz(@toyxyz3)_20230214-152406_1625516237794189312_photo.jpg



twitter_toyxyz(@toyxyz3)_20230214-152607_1625516745678286851_photo.jpg



twitter_toyxyz(@toyxyz3)_20230214-152650_1625516926893174784_photo.jpg



twitter_toyxyz(@toyxyz3)_20230214-152905_1625517492415401985_photo.jpg



twitter_toyxyz(@toyxyz3)_20230214-175359_1625553956389941248_photo.jpg







twitter_toyxyz(@toyxyz3)_20230216-191143_1626298297211326465_photo.jpg



twitter_toyxyz(@toyxyz3)_20230216-200903_1626312724379754496_photo.jpg



twitter_toyxyz(@toyxyz3)_20230216-202518_1626316812777586688_photo.jpg



twitter_toyxyz(@toyxyz3)_20230216-203401_1626319006306295808_photo.jpg



twitter_toyxyz(@toyxyz3)_20230216-211224_1626328668686684162_photo.jpg



twitter_toyxyz(@toyxyz3)_20230217-144723_1626594162060718083_photo.jpg






twitter_toyxyz(@toyxyz3)_20230218-160840_1626977005270102016_photo.jpg



twitter_toyxyz(@toyxyz3)_20230218-160917_1626977159184273408_photo.jpg



twitter_toyxyz(@toyxyz3)_20230219-072342_1627207280541437953_photo.jpg



twitter_toyxyz(@toyxyz3)_20230219-072344_1627207289722789889_photo.jpg



twitter_toyxyz(@toyxyz3)_20230219-154140_1627332598094761984_photo.jpg









twitter_toyxyz(@toyxyz3)_20230214-173040_1625548088856428544_photo.jpg



twitter_toyxyz(@toyxyz3)_20230131-174957_1620479514634760193_photo.jpg



twitter_toyxyz(@toyxyz3)_20230131-223426_1620551106857435136_photo.jpg
 
Last edited by a moderator:
tech moving fast

View attachment 4580245


View attachment 4580249


View attachment 4580261


View attachment 4580269


View attachment 4580277


View attachment 4580284


View attachment 4580297

View attachment 4580305


View attachment 4580324


View attachment 4580333


View attachment 4580340


View attachment 4580346


View attachment 4580358


View attachment 4580368


View attachment 4580378


View attachment 4580405


View attachment 4580420


View attachment 4580429



View attachment 4580446


View attachment 4580466


View attachment 4580473


View attachment 4580478


View attachment 4580482


View attachment 4580492


View attachment 4580496


View attachment 4580500


View attachment 4580508


View attachment 4580518


View attachment 4580542


View attachment 4580550


View attachment 4580561


View attachment 4580562


View attachment 4580565


View attachment 4580568


View attachment 4580569


View attachment 4580574
Is this 3d into mostly 2d art? The animation is the most impressive part. Is this some custom program toyxyz3 designed or is img2img in newer stable diffusion instances so configurable now it can recognize a skeleton and turn it into a fully detailed frame?
 
Is this 3d into mostly 2d art? The animation is the most impressive part. Is this some custom program toyxyz3 designed or is img2img in newer stable diffusion instances so configurable now it can recognize a skeleton and turn it into a fully detailed frame?
controlnet
it has a lot of different modes and it's basically like img2img on steroids
 
Even when the hands are fed in as a base model it still messes them up. Fascinating. AI really struggles with drawing fingers.
Image ai struggles with anything that isn't in a constant pose from what it's learning from. This applies to bodies, too. Especially fucks up if more than one person or animal is on screen in the same space. You want a single humanoid character standing in one of a few poses? Perfect. You want that same character holding something, touching someone else, being within a few inches of a piece of furniture and it starts to struggle a lot more. Not that it's impossible to get good images, and it's improving all the time but it is still a weak point. You can still see this in other areas like when the ai actually tries to make text itself, but as with that as well it's clearly getting better and fast. They went from producing gibberish to making somewhat comprehensible varieties of words in different fonts.
 
  • Agree
Reactions: A Hot Potato
"The US government, and all other governments, should regulate the development of SMI," he wrote, referring to superhuman machine intelligence. "In an ideal world, regulation would slow down the bad guys and speed up the good guys — it seems like what happens with the first SMI to be developed will be very important."

Altman's Sunday tweet about regulation echoed comments made by Mira Murati, the CTO of OpenAI, who said in a February 5 interview with Time magazine that ChatGPT should be regulated, as it could be misused. "It's not too early" to regulate it, Murati told the media outlet.
Just another example of OpenAI wanting to pull the ladder up.
 
Regarding controlnet, someone posted the pruned versions of all the networks it uses on civitai. Makes the whole thing a 6gb download instead of 60. It's easy to install too, just get the extension and throw the nets into its folder and you're done.

I decided to try the non-coomer route, and make a few Batman images. They're not perfect, but they are impressive.
width=400.jpg


Might have better results with this checkpoint :

A thing about diffusers is that they can't generate dark colors (especially solid black) well unless specifically shoved towards them in the training stage. This one is focused entirely on that. It does a great job with fake CG, too - Check out their example pdf and try copy pasting the environment prompt and changing the subject, it makes very authentic looking Resident Evil/Silent Hill backgrounds.
 
After several attempts at fixing my installation after an update broke compatibility with Windows 7, I opted to just do a full reinstall under Windows 10 with the latest build and grabbed xformers while I was at it because why not.

I have an older graphics card, a GTX 1080 (8 GB VRAM) and unlike before, VRAM is no longer the immediately limiting factor on how high a resolution I render an image. It's now how fast my card can render the image, and even then it's still significantly faster than it was prior. While rendering at 4k probably isn't in my reach (I haven't even tried), rendering at my monitor's native resolution is now entirely possible instead of upscaling to it instead.

I was expecting better performance, but not to this degree. I'm pretty impressed.
00028-88292636.jpg
 
tech moving fast

View attachment 4580245


View attachment 4580249


View attachment 4580261


View attachment 4580269


View attachment 4580277


View attachment 4580284


View attachment 4580297

View attachment 4580305


View attachment 4580324


View attachment 4580333


View attachment 4580340


View attachment 4580346


View attachment 4580358


View attachment 4580368


View attachment 4580378


View attachment 4580405


View attachment 4580420


View attachment 4580429



View attachment 4580446


View attachment 4580466


View attachment 4580473


View attachment 4580478


View attachment 4580482


View attachment 4580492


View attachment 4580496


View attachment 4580500


View attachment 4580508


View attachment 4580518


View attachment 4580542


View attachment 4580550


View attachment 4580561


View attachment 4580562


View attachment 4580565


View attachment 4580568


View attachment 4580569


View attachment 4580574
What really bothers me about Controlnet is the fact that you could do most of this with img2img and how lazy the community is getting while at the same time, doing vastly more work. In fact, I've been throwing daz models and my drawings into img2img since the whole thing kicked off and I don't really care about controlnet. I'm fucking psyched over Webui Two Shot, it completely solves the shared features issue when generating multiple people.

Don't get me wrong, LoRas are wonderful and so is controlnet, but now,on other ai generating forums, we're getting buried by people asking basic and simple questions.

I can't share the images due to how lewd they are, but I got into an argument with a coomposter and recreated his exact image, without using the several loras embedded in the prompt. his whole counter argument was " I can do it faster", my argument is " are you even doing anything at this point?" If the Lora brings back the same exact image/character, every single time, then why even bother ? You're just generating stuff that already exists due to how over/under trained most Loras are. Fuck. Just get some tracing paper or fire up photoshop at this point.


Anyway, now we're entering dangerous territory.

First worry.
People will completely forget how to prompt and will be crippled, relying on Loras and Controlnet to do things. There's less creativity now and more spoonfeeding. We were already drowning in all kinds of huge titty and giant monsters and scifi landcapes, but at that point in time, they were all different. Now we're getting bombarded with the same exact thing endlessly.
It's just trivial nonsense at this point. Every so often, there's a cute LoRa that does something like turn everything into Mashed Potatoes or render everything in Bob's Burger's style or adjusts the skies to a specific time and then there's the endless X character loras. There's so many same-face characters that it all seems pointless.

Now with the Lora thing, there's a kerfluffel with an overtrained Lora spitting out exactly what it was put in and producing the exact same words and copyrighted logos when mixed with controlnet. Giving the people who hate Ai easy access to actual weapons that can be successfully used against Ai is not a good idea. These guys are actually making a little headway with their done wrong examples, now that things are advancing, if they actually sat down and learned how to do it , they could actually be successful in their litigation.

It should remain just technical enough that you can't generate a perfect nike logo on every single attempt, but of course, retards want their infinitely same faced waifus to have the hottest kicks.


Second
We're quickly moving into the GIVE ME MONEY phase faster than anticipated, folks are really out here trying to sell merges. Civitai is looking like Lindenlabs/GaiaOnline shopfront circa 2009 right now. Thousands upon thousands of cpkts and loras, only a fraction actually do something different. Endless false advertising in the prompts as well. I've become a real party pooper on the site going around and bitching about how many fucking Loras and inversions are used to get said images. If I need 10+ inversions and three Loras to get something similar to what you're claiming this model can produce... What the fuck does your model actually do then?

Third worry
Automatic1111 is getting slower and more bloated with every new update. Other self run generation programs can't even img2img or run safetensors, once against fueling the retarded pigeonholing that's about to happen.

tldr; Everyone's putting all their eggs into one basket and everyone should have at least one commit/backup from before the torch update stored on an external HD along with the early versions of 1.4,1.5, Novel, Midjourney. I have very little trust in the ai community these days.
 
People will completely forget how to prompt and will be crippled, relying on Loras and Controlnet to do things. There's less creativity now and more spoonfeeding. We were already drowning in all kinds of huge titty and giant monsters and scifi landcapes, but at that point in time, they were all different. Now we're getting bombarded with the same exact thing endlessly.
It's just trivial nonsense at this point. Every so often, there's a cute LoRa that does something like turn everything into Mashed Potatoes or render everything in Bob's Burger's style or adjusts the skies to a specific time and then there's the endless X character loras. There's so many same-face characters that it all seems pointless.
I made the mistake of taking a glance at DeviantArt's AI art section. It's just endlessly repeating renders of the same fat, naked character using a rather bland prompt/model that looks a lot like the NovelAI model from the leak back in August. And of those that aren't chubby chasers, many of the uploads are of images that have some distinct flaw that easily marks them as AI generated as the creator can't be bothered to take a few minutes in Photoshop to fix things up to make a better image.
Automatic1111 is getting slower and more bloated with every new update. Other self run generation programs can't even img2img or run safetensors, once against fueling the retarded pigeonholing that's about to happen.

tldr; Everyone's putting all their eggs into one basket and everyone should have at least one commit/backup from before the torch update stored on an external HD along with the early versions of 1.4,1.5, Novel, Midjourney. I have very little trust in the ai community these days.
After my latest kerfuffle with updating, I'm no longer going to be updating existing installations anymore. Maybe once a month, I'll do a fresh install with the updates but I refuse to change anything with the older version that's working perfectly well.
 
One theory I had: if AI gets so good it makes art preferable to actual people, then the medium would change from a few big creators/producers making mainstream series everyone's seen, to instead a community-based thing where you join a specific community of curators looking for something actually good the AI created to consume next. Instead of watching/reading the next big thing, you'd join a niche discord or whatever of people sharing AI-created stories they've seen that turned out actually good and appealing for that specific community you joined.
 
One theory I had: if AI gets so good it makes art preferable to actual people, then the medium would change from a few big creators/producers making mainstream series everyone's seen, to instead a community-based thing where you join a specific community of curators looking for something actually good the AI created to consume next. Instead of watching/reading the next big thing, you'd join a niche discord or whatever of people sharing AI-created stories they've seen that turned out actually good and appealing for that specific community you joined.
Or find actually creative ways to be ugly. I'm pretty sure you could have a model to spit out soulless Corporate Memphis garbage all day long.
 
After several attempts at fixing my installation after an update broke compatibility with Windows 7, I opted to just do a full reinstall under Windows 10 with the latest build and grabbed xformers while I was at it because why not.

I have an older graphics card, a GTX 1080 (8 GB VRAM) and unlike before, VRAM is no longer the immediately limiting factor on how high a resolution I render an image. It's now how fast my card can render the image, and even then it's still significantly faster than it was prior. While rendering at 4k probably isn't in my reach (I haven't even tried), rendering at my monitor's native resolution is now entirely possible instead of upscaling to it instead.

I was expecting better performance, but not to this degree. I'm pretty impressed.
View attachment 4610924
Prompt?
 
modelshoot style, (extremely detailed CG unity 8k wallpaper), full shot body photo of the most beautiful artwork in the world, cat spirit, (cat ears:1.3), elegant, mythical, mystical, (fantasy:1.5), professional majestic oil painting by Ed Blinkey, Atey Ghailan, Studio Ghibli, by Jeremy Mann, Greg Manchess, Antonio Moro, trending on ArtStation, trending on CGSociety, Intricate, High Detail, Sharp focus, dramatic, photorealistic painting art by midjourney and greg rutkowski
Negative prompt: canvas frame, cartoon, 3d, ((disfigured)), ((bad art)), ((deformed)),((extra limbs)),((close up)),((b&w)), wierd colors, blurry, (((duplicate))), ((morbid)), ((mutilated)), [out of frame], extra fingers, mutated hands, ((poorly drawn hands)), ((poorly drawn face)), (((mutation))), (((deformed))), ((ugly)), blurry, ((bad anatomy)), (((bad proportions))), ((extra limbs)), cloned face, (((disfigured))), out of frame, ugly, extra limbs, (bad anatomy), gross proportions, (malformed limbs), ((missing arms)), ((missing legs)), (((extra arms))), (((extra legs))), mutated hands, (fused fingers), (too many fingers), (((long neck))), Photoshop, video game, ugly, tiling, poorly drawn hands, poorly drawn feet, poorly drawn face, out of frame, mutation, mutated, extra limbs, extra legs, extra arms, disfigured, deformed, cross-eye, body out of frame, blurry, bad art, bad anatomy, 3d render, text, artist, credit, description, (extra ears)
Steps: 30, Sampler: DPM++ 2M Karras, CFG scale: 10, Seed: 88292636, Size: 768x432, Model hash: 08acb74861, Model: dreamshaper_33, Denoising strength: 0.45, Hires upscale: 2.5, Hires steps: 60, Hires upscaler: SwinIR_4x
 
Am I the only one who noticed that the hands and fingers in AI generated pictures look better now in comparison to when it all started?

I'm using Artbreeder and it's incredible how much the quality improved.

also @Pendulous Dong

Absolutely beautiful pictures! Love the one with the bird especially.
I would agree if I wasn't getting 504 Gateway Time-Out errors when I try to make an account on Artbreeder.
Back when it worked for me, yeah, it was a blast using Artbreeder.
 
Fwiw, if anyone else has a 30xx or 40xx series card:

E: DISCLAIMER: I've been fucking around with this for like two hours now and it's erratic and seems to break some models. Back up your DLLs (everything in /stable-diffusion-webui/venv/lib/site-packages/torch/lib/) if you're going to try this.


Grab whatever version matches your system, open it, find /bin/ in the archive, extract all the DLLs there to /stable-diffusion-webui/venv/lib/site-packages/torch/lib/ replacing the ones that were there already.

Enjoy your absolutely absurd speedup. On a 4070ti I'm getting 1080p 70-steps generations in 0:49 and standard 512x512 30-steps in 0:02. For comparison those times were around 1:35 and 0:06 before.

ETA: Something else that's far more entertaining than it should be: If you just use a static wallpaper on your computer, throw the image into the deepbooru interrogator to get a prompt back for it (or this, which is trickier but more reliable) and just feed it back into img2img with that prompt, batchsize one, high batch count to get a bunch of very very similar but not identical images. Throw all the results into a folder, set it to slideshow shuffle every couple hours.
 
Last edited:
Back